Summary: Perplexity launched Model Council on February 5, 2026 to send one query to three frontier models and synthesize their agreements and disagreements with another model. The design makes cross-model verification visible in search, while its multiple inference runs confine it to consequential questions where reliability justifies higher cost and latency.
Perplexity announced Model Council in a February 5, 2026 post on its official blog, and the changelog published a day later grouped it with an upgraded Deep Research, a simplified input bar, step-by-step learning, and memory improvements as that week’s combined release. The design compresses into one sentence: one question goes to three frontier models at once, a separate model reviews the three responses, and the result is distilled into a single answer — agreements and disagreements both flagged.
This is not another chat skin. It is an attempt to take multi-model cross-verification, historically a research method, and ship it as a working feature of a search product. As trust in a single model’s output keeps eroding, Perplexity moved the verification work inside the product itself.
The Mechanics of Model Council
A February 9 report from Livemint sketched the flow: a user’s query is dispatched to three models simultaneously, the system compares the three responses, and surfaces key insights alongside the points where they diverge, letting users gauge answer reliability. The output, in other words, is no longer “what one model said” but a structured account of what each model said, where they line up, and where they fight.
On hard questions, disagreement carries information of its own. When three models return different answers to the same question, that gap marks exactly the spot where human judgment or deeper checking belongs. Bury it, and the user walks away with one confident wrong answer.
The Rest of That Week’s Release: A Council, and More
The February 6 changelog shows Model Council was one item in a broader refresh: Deep Research received an upgrade, the input bar was streamlined, step-by-step learning arrived, and memory got better. The timing is also telling — this was the week frontier-model competition turned white-hot, with Anthropic releasing Claude Opus 4.6 on February 6. “Which model is best” gets re-answered every week, and Perplexity’s reply is not to choose a winner on the user’s behalf but to turn the single-choice question into a side-by-side comparison.
The Reason Councils Work: Mistakes That Rarely Coincide
Multi-model ensembling rests on established research: models trained independently don’t fail in identical ways. The odds of three models committing the same random error at the same time sit far below the odds of one model erring alone, so comparison cancels noise — surviving disagreement tends to mark the genuinely difficult or uncertain points.
Model Council’s key contribution as a product is that it refuses to pretend disagreement away. Conventional answer synthesis smooths the differences into one blended version; this design marks them instead and leaves the call to the user. On factual questions, skim the paragraphs where all three models agree; the ones worth a careful read are where they argue.
The Price: Three Inference Runs Plus One Synthesis
Each question means three frontier models plus one review-and-synthesis pass, so latency and cost run to multiples of a single-model pipeline. That constraint defines the use case: everyday queries can stay on a single model, while consequential decisions — medical, legal, financial, fact-checking — justify convening the council. For developers this is the textbook accuracy-for-cost trade, and it explains why comparable features in the market usually open first on premium tiers: only users willing to pay for reliability can support a triple inference bill.
The Takeaway for Product Teams: Routing and Councils Are Two Different Architectures
A multi-model strategy forks into two routes. Routing sizes up the question first and forwards it to the single best-fit model — it buys savings. A council puts several models in mutual check — it buys dependability. The former optimizes marginal cost; the latter, how far the answer can be trusted. As capability gaps between frontier models keep narrowing, “can the answer be verified” is turning into the differentiator itself — Perplexity wasn’t the first team to think of ensembling, but it ranks among the earliest vendors to put one in a mass-market search box. For product teams designing AI features right now, the question shifts from “which model do we use” to “which questions justify triple the cost.”